跳到导航

Tools as code

ElevenAgents tools can now be built in code with the tool builder in the Python and JavaScript SDKs, from v2.79.0. Define webhook, client and system tools with chained calls such as get(), header() and assign(), then pass the result of build() to tools.create() or tools.update(). The builder keeps path parameters, required lists and value sources consistent, runs entirely in the SDK, and produces tools you can still edit in the dashboard. See Tools as code.

Lower Speech to Text prices

Speech to Text now costs $0.20 per hour of audio for batch transcription and $0.22 per hour for realtime transcription, for every Scribe model on all self-serve and Business plans. Enterprise pricing is unchanged. See Speech to Text for the available models.

Content Provenance API

Content provenance has a beta endpoint, POST /v1/content-provenance/watermark, that checks an uploaded audio file for an ElevenLabs watermark. It returns is_detected and, when the generation can be identified, the product, model, generated_at time and unique_id of the generation, never the prompt or the account behind it. API keys need the new watermark permission. See Content provenance.

Build and publish agents from the hosted MCP

Hosted MCP adds 39 tools for building ElevenAgents from any MCP client, covering workflow edits, attaching and detaching tools, test runs, agent generation, phone number import, batch calling, workspace secrets and ElevenAgents settings. Edits are staged in a draft, and only the publish tool makes them live. Connected clients ask for consent once more because of two new read scopes. See Hosted MCP.

All changes

  • Model deprecations: gpt-4.1, gpt-4.1-mini, gpt-4o and gpt-4o-mini are deprecated ahead of their shutdown on January 14, 2027. Requests fall back to gpt-6-luna, for 10% of requests from December 15, 2026 and for all requests from January 9, 2027. The deprecated gpt-5, gpt-5-mini and gpt-5-nano now fall back to gpt-6.1-sol, gpt-6.1-terra and gpt-6-luna instead of the GPT-5.6 models announced on November 2. The gemini-3.6-flash fallback to gemini-3.8-flash reaches all requests on November 17, ahead of its retirement on November 19. Models past their provider shutdown dates, including Claude 3.5 Sonnet, Claude 3.7 Sonnet, Claude 3 Haiku, Claude Sonnet 4, Gemini 2.0 Flash, Gemini 2.0 Flash-Lite, gpt-4-turbo and gpt-3.5-turbo, resolve to their replacement models immediately, and grok-beta is removed from the LLM enum. See Models.
  • Fast tier for Gemini: Gemini agent models accept service_tier: fast at priority rates, and a request is billed at the fast rate only when it is served at priority. Models that were already served at priority are billed at the standard rate until December 12. Custom guardrails add service_tier (default standard) and a matching picker in the dashboard; Gemini priority processing for guardrails stays free until December 4, after which only guardrails set to fast use it and are charged for it. gpt-6-luna agents without a reasoning effort setting now run at low instead of medium. The custom guardrail model picker no longer offers Gemini 2.5 Flash and Gemini 2.5 Flash Lite; existing guardrails and the API still accept them.
  • Knowledge base: Get chunking override and Update chunking override set max_chunk_size, overlap_percentage and custom_separators for a document or folder, with null inheriting from the parent folder, and the setting is available for every document and folder in the dashboard. Discover now on auto-discover crawl folders is available to every workspace. Folder responses return crawl_job_id for every crawl root folder. Voice agents read Markdown links in documents as their link text, and documents scraped from URLs keep ordered-list numbers without long runs of empty lines. The deprecated dependent_agents field in knowledge base list and search responses is always empty. See Knowledge base.
  • Conversation history: List conversations and message text search filter by tool ID with tool_ids, tool_ids_successful and tool_ids_errored, so tools that share a name are told apart, and by number with phone_number_ids and whatsapp_phone_number_ids. Tool results in transcripts record tool_id, and function-backed integration tool calls return function_details with the parameters the LLM sent. Get conversation file lets server tools download a file the user attached during a live conversation, and the system__uploaded_file_ids dynamic variable passes the conversation’s attachment IDs to tool parameters.
  • Workflows and procedures: Workflow sub-agent nodes add allow_interruptions_on_entry (default false) and an Allow interruptions switch, so callers can interrupt a node’s first response when it is set to generate immediately. Start procedure nodes are available to every workspace. Structured procedures add the update_state system tool, which sets dynamic variables, and procedure responses list written_dynamic_variables. The transfer_to_number system tool accepts transfer_rules, and workflow phone number nodes accept provider_configs, with one transfer configuration per telephony provider. Dead end analysis no longer flags nodes whose conditions cover every case. See Structured procedures.
  • Realtime API: response.done includes usage, and tool call responses report the usage of the LLM call that made the tool call. conversation.item.create rejects malformed items, and every error for a rejected client event echoes its event_id. Realtime agents honor tts.expressive_mode, and the Expressive mode switch is shown for them. The LLM usage calculator returns prices for eleven_realtime_v1_mini and eleven_realtime_v1_max. See Realtime models.
  • Branches: The Merge branch and Pull in changes dialogs resolve conflicts field by field for every workspace, replacing the read-only conflict list and the Force apply switch. Merges and rebases without conflict_resolutions compare voices, tools and tests entry by entry, as merge previews do. The Environment variables tab in Version Control is renamed Environments and opens on a view of how this agent’s variables differ by environment.
  • Privacy: Conversation history redaction adds the criminal_record entity type, redacts multi-paragraph call summaries and titles, and masks entities next to Korean particles and in Chinese and Japanese text without spaces. On PCI-compliant workspaces, PATCH /v1/convai/settings rejects pre-call and post-call webhooks that are not on the allowlist with 400 pci_compliance_violation. See Conversation history redaction.
  • Channels and telephony: Voximplant calls can stream to an agent over a single WebSocket from a VoxEngine scenario; the integration is in limited availability and enabled per workspace. The WhatsApp integration adds a whatsapp_send_image tool that sends a JPEG or PNG image of up to 5 MB from a public URL, with an optional caption of up to 1,024 characters. In a Genesys Open Messaging handover, the agent tells the customer when a human agent joins.
  • Dashboard and analytics: Architect can ask clarifying questions with single- or multi-select options, close merge proposals and report who has access to an agent. Agent routes check only the permissions they use, so custom roles get the access they were granted, and viewers can no longer generate workflows or start batch calls. Tool failure alerts are tracked by tool ID and count code tool failures. Turn-taking latency is capped at the time since the agent finished speaking. Data collection discards string values outside the field’s enum. Conversation billing starts when the WebSocket is accepted after setup.
  • Fixes: Interrupted turns no longer send stale retries to custom LLMs. A mid-turn fallback to another LLM no longer repeats the reply or runs words together. A structured procedure that ends a call after a sub-procedure step no longer fails with Unsupported node type. A guardrail that fails on its final retry no longer leaves the blocked message in the transcript. Editing an auth connection takes effect on the next token request instead of after the cached token expires. Tool dependent agents include branches that use the tool only on workflow nodes.
  • Lower prices: Every Scribe model costs $0.20 per hour for batch transcription and $0.22 per hour for realtime transcription on self-serve and Business plans. Enterprise pricing is unchanged.
  • Language detection in partials: With include_language_detection=true, realtime partial_transcript messages include language_code once at least 4 words are detected with confidence on 2 consecutive partials. language_code on committed transcripts with timestamps can be null.
  • Fixes: Realtime transcription no longer sends transcripts that contain only punctuation, such as ..., during silence, or a partial transcript after the committed transcript for the same segment. Committed transcripts with timestamps capitalize the first letter like text-only commits. Speech after a long continuous stream of about 36 seconds is no longer lost. Text frames that are not valid JSON return input_error and keep the session open instead of closing it with internal_error.
  • New tools: 39 tools cover agent draft edits, publishing, test runs and simulation tests, agent generation, phone number import, batch calling, workspace secrets, ElevenAgents settings, signed URLs and WhatsApp. Further tools read account usage, conversation audio, pronunciation dictionaries, workspace webhooks without their secrets, ElevenAgents concurrency limits and auto top-up state, and query conversations across every agent you can view. Agents can use Voice Library voices. See Hosted MCP.
  • New scopes: MCP clients request the user_read and pronunciation_dictionaries_read scopes, so each connected client asks for consent once more.
  • Voice listing: creative_list_voices pages its results with page_size (up to 100) and cursor/next_cursor instead of returning only the first 25 voices.
  • Fixes: Tools no longer ask the client for a context argument describing the user’s goal. Validation errors return isError with the public tool name.
  • Webhook testing and redelivery: Send test webhook event sends a signed sample event once and returns the endpoint’s status code and response time. Workspace admins can use Redeliver webhook delivery from a web session to send a stored delivery again with the same delivery ID and an ElevenLabs-Webhook-Redelivery: true header; repeating it within 5 seconds returns 409 redelivery_in_flight. Test events and redeliveries share a limit of 10 per webhook per minute.
  • Agents environments permission: The agents_environment_variables_write permission, shown as Manage Agents environments in group settings, controls creating, updating and deleting agent environments and environment variables. Admins with Agents write access keep their access.
  • SCIM: Users can belong to several seat-type groups and receive the highest seat type among them. With SCIM routing on, PATCH and PUT on a listed workspace group convert it to a routing group instead of returning Group not found.
  • Audit logs: OAuth app authorizations, management and revocations are recorded, with three new AuditActivityName values.
  • Watermark check: POST /v1/content-provenance/watermark (beta) checks an uploaded audio file for an ElevenLabs watermark and returns is_detected and details with product, model, generated_at and unique_id. Send version to choose the decoder; only latest is supported. It requires the watermark permission and does not accept OAuth access tokens. See Content provenance.
  • Voice design permission: POST /v1/text-to-voice/design, POST /v1/text-to-voice, POST /v1/text-to-voice/create-previews and POST /v1/text-to-voice/{voice_id}/remix require the voice_generation permission on API keys, OAuth grants and workspace roles, and return 401 missing_permissions without it.
  • Voice Changer sharing: Voice Changer history items can be shared with workspace members from the share dialog.
  • Fixes: GET /v1/text-to-voice/{generated_voice_id}/stream returns 404 voice_not_found for a preview that does not exist instead of 500. Editing a Professional Voice Clone’s name, description or labels no longer retrains the voice.
  • Deprecated parameter: apply_language_text_normalization on the Text to Speech endpoints is deprecated.
  • WebSocket errors: The stream-input and multi-stream-input WebSockets close with 1008 model_not_found for an unknown model_id instead of 1011.
  • Fixes: Text edited on a v2 model is kept when switching back to a multi-speaker Eleven v3 or Eleven v4 dialogue.
  • Pronunciation highlights: Audiobook projects highlight words matched by applied pronunciation rules, and a dialog breaks down conflicts when several rules match.
  • Fixes: Duplicated or trimmed clips of a long video no longer freeze timeline playback. Video exports whose source decoder stalls, and exports of H.264 sources that switch parameter sets, fail with an error instead of hanging.
  • Flows preferences: Flows has preferences for default image, video, LLM and speech models for new nodes, video autoplay, connection visibility, notification sounds and snapping guides.
  • Image editor: Editing sessions are saved to the server and resume after closing the tab or on another device, including while an AI operation runs. The Brand kits tab shows the active kit’s colors, typefaces, logos and references, which can be dragged onto the canvas, and the Layout panel switches between percent and pixel units.
  • Delete takes: Each take on the History page has a delete button.
  • Immediate cancellation: GET /v1/user/subscription returns cancels_immediately (boolean, default false). Self-serve plans whose renewal is waiting on 3-D Secure authentication now cancel right away and void the unpaid invoice instead of cancelling at the end of the period.
  • Reception API: Reception.ai partners can manage their customers’ workspaces in code with 75 endpoints under /v1/reception. They can create a workspace for each customer, sell it one of their plans, follow billing, and set up the workspace’s business profile, receptionist (with rules, procedures and transfer rules), phone numbers, services, staff, locations, holidays, products, assets, knowledge sources and team. The endpoints are in the Python and JavaScript SDKs from v2.79.0 as reception. Requests need an API key created in the reseller admin panel; other API keys are rejected with not_a_reseller_api_key. See Reception API and the quickstart.

JavaScript SDK

  • v2.78.0 - Speech Engine streams keep the answer text from Gemini models instead of dropping it.
  • v2.79.0 - Added a tool builder (Tool, Schema and Header, exported from the package root) that defines webhook, client and system tools with chained calls and produces the request for tools.create() and tools.update(). See Tools as code. Added the reception client for the Reception API.

Python SDK

  • v2.78.0 - Speech Engine streams keep the answer text from Gemini models instead of dropping it.
  • v2.79.0 - Added the tool builder in elevenlabs.agent_tools, which defines webhook, client and system tools with chained calls and produces the request for tools.create() and tools.update(). See Tools as code. Conversation and AsyncConversation accept callback_unhandled_client_tool_call, called when the agent requests a client tool that has no registered handler. Added the reception client for the Reception API.

ElevenLabs CLI

  • v1.9.2 and v1.9.3 - Publishes the v1.9.1 changes to npm, Homebrew and Scoop, which the v1.9.1 release did not reach. The installed say skill uses a current default voice and notes that free plans can use premade voices but not Voice Library voices.
  • v1.10.0 - onboard treats a page that embeds the agent widget as wired, and ends a voice agent project with a link to talk to the agent and its dashboard page.
  • v1.11.0 - Agent, tool and test configs can be .json, .yaml/.yml or .toml files, chosen by file extension. agents add and agents pull take --config-format to set the format of new files.
  • v1.12.0 - Updated the bundled API client.
  • v1.12.1 - tools diagram draws one tool as a tree, in the terminal or as Mermaid, with each parameter’s value source, headers, auth connection and response handling. Workflow diagrams label tool nodes with names from tools.json. agents pull --sparse keeps existing minimal configs minimal instead of writing every default.

Packages

New endpoints

ElevenAgents

  • Get conversation file - GET /v1/convai/conversations/{conversation_id}/files/{file_id} downloads a file attached to a conversation.
  • Get chunking override - GET /v1/convai/knowledge-base/{documentation_id}/chunking-override
  • Update chunking override - PATCH /v1/convai/knowledge-base/{documentation_id}/chunking-override with chunking_override (max_chunk_size, overlap_percentage and custom_separators, each a nullable integer or string array).

Workspaces

Reception.ai

  • Reception.ai - 75 endpoints under /v1/reception for Reception.ai partners. Reseller-level groups are workspaces, tiers, subscriptions and billing. Workspace-level groups, which take the xi-workspace-id header, are business_profile, receptionists (with rules, procedures and transfer_rules), phone_numbers, services, staff, locations, holidays, products (with stock), assets, knowledge_sources and team (members and invites).

Content provenance

  • Check watermark - POST /v1/content-provenance/watermark (beta) accepts a multipart file and optional version.

Updated endpoints

ElevenAgents

  • List conversations
    • Added optional tool_ids, tool_ids_successful, tool_ids_errored, phone_number_ids and whatsapp_phone_number_ids query parameters (string arrays). Message text search (GET /v1/convai/conversations/messages/text-search) accepts the same five.
    • Added optional data_collection_filter and dynamic_variable_filter query parameters (string), single-value forms of data_collection_params and dynamic_variable_params.
  • Get conversation details, test and simulation responses
    • Tool results in transcripts add nullable tool_id (string).
    • Integration webhook tool calls add nullable function_details.
  • Create agent, Update agent and the agent get endpoints
    • TransferToNumberToolConfig adds transfer_rules (array of TransferRule, each with a condition and provider_configs for Twilio, SIP trunk or Exotel).
    • Workflow phone number nodes add provider_configs (array of Twilio, SIP trunk or Exotel transfers), mutually exclusive with a non-empty transfer_destination.
    • Workflow sub-agent (override_agent) nodes add allow_interruptions_on_entry (boolean, default false).
    • CustomGuardrailConfig adds service_tier (LLMServiceTier, default standard).
    • ConfigEntityType adds criminal_record.
    • Procedure references add written_dynamic_variables (string array), also on List procedures.
    • The LLM enum, also used by conversation overrides on the /v1/convai/conversation WebSocket, removes grok-beta.
  • Knowledge base document and summary responses add nullable chunking_override. dependent_agents on summary responses is deprecated and always empty.

Text to Speech

Speech to Text

  • Realtime WebSocket - PartialTranscript adds nullable language_code (string). language_code on CommittedTranscriptWithTimestamps and on SessionStarted.config is nullable.

Workspaces

  • Service account API keys and workspace groups - PermissionType and WorkspaceGroupPermission add agents_environment_variables_write and watermark.

Billing